Dependency Forest for Medical Relation Extraction

Code corresponding to our paper "Leveraging Dependency Forest for Neural Medical Relation Extraction" at EMNLP 2019

Introduction

Folder "biaffine_forest" is a deep biaffine parser that supports producing dependency forests as introduced in our paper. It is obtained by adapting the original dozat-parser. To generate forests, use "--nbest" or "--cubesparse" instead of the traditional "--test" option when decoding. "--nbest" and "--cubesparse" correspond to our "KbestEisner" and "Edgewise" method, respectively. Parser training remains the same as the original system. As mentioned by the original system description, this parser only works on Python 2 and TF 0.1.2, a very old version.

Folder "re_forest_grn" is our main relation extraction (RE) system based on graph recurrent network (GRN) for consuming dependency trees/forests. There are several scripts within the "re_forest_grn/data" folder to help generating necessary data, such as word embeddings. Training and decoding shells are also provided. This model is based on Python 2 and TF 1.8.0.

Data

Since the orignal website for data obtaining is not available, you may download the data through CPR and PGR. I also attach my script for preprocessing. It uses several handcrafted rules to pre-tokenize some special entities, which cannot be tokenized by a standard tokenizer. But, we just realized recently that this could be done in a smarter way that simply pre-tokenizes with the already provided character-based positions. You may improve this tokenizer for better performances.

Cite

If our work helps your research/system, please cite our work with the following bibtex file.

@inproceedings{song-etal-2019-leveraging,
    title = "Leveraging Dependency Forest for Neural Medical Relation Extraction",
    author = "Song, Linfeng  and
      Zhang, Yue  and
      Gildea, Daniel  and
      Yu, Mo  and
      Wang, Zhiguo  and
      Su, Jinsong",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
    year = "2019",
    doi = "10.18653/v1/D19-1020",
    pages = "208--218"
}